Abstract
Steady-State Visual Evoked Potentials (SSVEPs) offer a robust basis for brain–computer interface (BCI) systems due to their high signal-to-noise ratio, minimal user training requirements, and suitability for real-time decoding. In this work, we propose MultiHeadEEGModelCLS, a novel Transformer-based architecture that integrates context-aware representation learning into SSVEP decoding. The model employs a dual-stream spatio-temporal encoder to process both the input EEG trial and a contextual signal (e.g., template or reference trial), enhanced by a learnable classification ([CLS]) token. Through self-attention and cross-attention mechanisms, the model aligns trial-level representations with contextual cues. The architecture supports multi-task learning via signal reconstruction and context-informed classification heads. Evaluation on benchmark datasets (Speller and BETA) demonstrates state-of-the-art performance, particularly under limited data and short time window scenarios, achieving higher classification accuracy and information transfer rates (ITR) compared to existing deep learning methods such as the multi-branch CNN (ConvDNN). Our method achieved an ITR of 283 bits/min and 222 bits/min for the Speller and BETA datasets, and a ConvDNN of 238 bits/min and 181 bits/min. These results highlight the effectiveness of contextual modeling in enhancing the robustness and efficiency of SSVEP-based BCIs.